rplksr DySample bug #296
Description
https://github.com/chaiNNer-org/spandrel/blob/main/libs/spandrel/spandrel/architectures/PLKSR/__arch/RealPLKSR.py#L147
This is not your fault, but rather a feature of the detection.
You will never get the condition DySample: True plus 1x in models from neosr so the condition with 1x is useless, also due to an error with groups 3x models do not work, it is better to replace this block with a piece from trainner-redux which we have with the-database changed. this is fully compatible with neosr but also extends the capabilities of DySample and makes spandrel compatible with redux.
https://github.com/the-database/traiNNer-redux/blob/master/traiNNer/archs/realplksr_arch.py#L142
if dysample:
groups = out_ch if upscaling_factor % 2 != 0 else 4
self.to_img = DySample(
in_ch * upscaling_factor**2,
out_ch,
upscaling_factor,
groups=groups,
end_convolution=True if upscaling_factor != 1 else False,
)
else:
self.to_img = nn.PixelShuffle(upscaling_factor)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.feats(x) + self.repeat_op(x)
return self.to_img(x)